DeepSeek V4 Flash Official API (0731): 82.7 Terminal-Bench, $0.14/$0.28 & Agent Upgrades

DeepSeek put the official DeepSeek-V4-Flash API into public beta on July 31, 2026. The checkpoint is Flash-0731. Call it with model ID deepseek-v4-flash. Per DeepSeek’s changelog, this is a re-post-train of the Flash preview: same architecture and size, substantially stronger agent results, native Responses API support, and first-class Codex integration.

This is a news update on top of the existing DeepSeek V4 coding guide. That guide covers the April 24 V4 family launch. Today’s post is only what changed with Flash-0731.

What the Flash-0731 announcement means

Per DeepSeek’s API change log (2026-07-31):

1. Official Flash API, public beta. Calling method is unchanged: set model to deepseek-v4-flash. The docs list the model version as DeepSeek-V4-Flash-0731.
2. Agent focus. DeepSeek says agent capabilities are “significantly enhanced,” with published agent-suite scores “far exceeding V4-Pro-Preview.”
3. Responses API + Codex. Flash natively supports the Responses API format and is “specifically adapted for Codex.” DeepSeek’s Codex integration guide states that only deepseek-v4-flash supports Codex today; Pro is expected in early August 2026.
4. Scope is Flash-only. DeepSeek is explicit: this update upgrades the V4-Flash API. The V4-Pro API and APP/WEB models are unchanged. The official V4-Pro release “will follow soon.”

Open weights for the checkpoint are on Hugging Face as deepseek-ai/DeepSeek-V4-Flash-0731 (MIT license on the model card; lastModified 2026-07-31). For router-style access, OpenRouter already lists deepseek/deepseek-v4-flash and a Flash-0731-dated variant — third-party routing, not the official API.

If you were waiting for a “real” Flash after the April preview: this is that cut, aimed at agent loops rather than chat demos.

Flash-0731 vs V4-Pro-Preview: the agent benchmark jump

DeepSeek’s published agent numbers (changelog + HF model card) put Flash-0731 ahead of both Flash Preview and V4-Pro-Preview on the agent suite they report:

Benchmark (per DeepSeek)	Flash-0731	Flash Preview	V4-Pro-Preview
Terminal Bench 2.1	82.7	61.8	72.1
NL2Repo	54.2	39.4	38.5
Cybergym	76.7	38.7	52.7
DeepSWE	54.4	7.3	12.8
Toolathlon-Verified	70.3	49.7	55.9
Agents’ Last Exam	25.2	15.8	16.5
AutomationBench Public	25.1	10.8	12.8
DSBench-FullStack †	68.7	37.0	41.8
DSBench-Hard †	59.6	25.8	31.1

† DeepSeek labels DSBench-FullStack / DSBench-Hard as internal sets.

Terminal-Bench 2.1 at 82.7 is the headline. That is +20.9 over Flash Preview (61.8) and +10.6 over V4-Pro-Preview (72.1), per the HF comparison table. DeepSeek also lists Flash-0731 against third-party models on the same card (e.g. Opus-4.8 at 85.0, GLM-5.2 at 81.0 on Terminal Bench 2.1) — vendor-reported rows, not independent re-runs.

How they measured agent tasks (per DeepSeek notes on both the changelog and HF card): public Code Agent benchmarks used DeepSeek Harness minimal mode (framework “to be released”), max effort, temperature=1.0, top_p=0.95.

Read the jump as agent post-training, not a new base model. DeepSeek states Flash-0731 keeps the same architecture and size as Flash Preview and was “only re-post-trained.” For coding agents that spend tokens on tools, terminal loops, and multi-step repos, that is the product change that matters.

Caveat, same as any vendor bench table: your repository, harness, and approval policy are the real gate. Treat 82.7 as a ranking signal for agentic terminal work, not a guarantee that Flash will land every PR cleanly.

Flash-0731 pricing and open weights

Official API rates from DeepSeek’s Models & Pricing page (USD per 1M tokens):

Item	deepseek-v4-flash	deepseek-v4-pro
Input (cache miss)	$0.14	$0.435
Input (cache hit)	$0.0028	$0.003625
Output	$0.28	$0.87
Context length	1M	1M
Max output	384K	384K
Concurrency limit (per docs)	2500	500
Responses API	Yes	Not yet (early Aug 2026 per docs)

So the Flash price cluster is real and official: $0.14 / $0.28 input miss / output, with cache-hit input at $0.0028. DeepSeek also notes a planned peak/off-peak policy (2× during Beijing peak windows) with an effective date still subject to official announcement — check the pricing page before you model long-run cost.

Open weights. The HF card for DeepSeek-V4-Flash-0731 lists MIT license. The April 2026 DeepSeek V4 model card PDF lists Flash as 285B total parameters, 13B activated per token, MoE, 1M context. Some community listings (e.g. OpenRouter’s Flash page) round total params to 284B; the official figure is 285B / 13B. Flash-0731 did not change architecture/size relative to Flash Preview, per DeepSeek.

OpenRouter. Searching “deepseek v4 openrouter” still maps to real routes: OpenRouter hosts deepseek/deepseek-v4-flash (preview-era listing) and a dated Flash-0731-style slug in its model catalog. Third-party prices differ from DeepSeek’s official $0.14/$0.28 (OpenRouter’s Flash page currently shows blended provider rates, often lower). For agent + Codex work, prefer the official API path DeepSeek documents; use OpenRouter when you want multi-provider routing, not when you need DeepSeek’s Responses/Codex setup verbatim.

Where V4-Pro stands

DeepSeek V4 Pro is not the official GA release yet. The API and APP/WEB models are unchanged since April.

What is true today, from fetched DeepSeek docs:

• The API has accepted deepseek-v4-pro since the April 24, 2026 V4 launch (preview-era API access).
• Pricing for deepseek-v4-pro is published ($0.435 / $0.87 cache-miss input / output).
• On July 31, DeepSeek said this update only upgrades Flash. “The DeepSeek-V4-Pro API and the APP/WEB models are unchanged.”
• DeepSeek’s wording on Pro’s official status: “The official release of DeepSeek-V4-Pro will follow soon.”
• Responses API and Codex: Flash yes now; Pro “expected… early August 2026” for Codex / Responses support per pricing footnotes and the Codex guide.

Pro is not “launched today”; the update is Flash-only. And Flash-0731 beats V4-Pro-Preview on DeepSeek’s agent suite (Terminal Bench 2.1 included), at roughly one-third the output token price. When the official Pro cut ships, re-check the same agent rows — until then, Flash-0731 is the agent-oriented DeepSeek V4 you can run with current Codex + Responses support.

For family-level Pro vs Flash trade-offs (non-news context), see the DeepSeek V4 coding guide.

Run DeepSeek V4 Flash with SeaWork

SeaWork remote-controls coding agents (Claude Code and Codex today) from your phone. Your code, credentials, and MCP servers stay on your machine; the control plane is what moves to your pocket. DeepSeek’s Flash-0731 announcement is explicitly about Responses API + Codex — the “agent running on a machine you may leave” story. SeaWork does not need to be the model gateway for that to matter.

SeaWork’s built-in gateway catalog does not need to serve deepseek-v4-flash for this to matter. Point your Codex (or Claude Code) provider at DeepSeek’s official API (https://api.deepseek.com, model deepseek-v4-flash — see DeepSeek’s Codex setup), then control that local agent with SeaWork.

seawork run --provider codex --model deepseek-v4-flash "your task"
seawork ls -a
seawork logs <id> -f

From the phone: watch every tool call live, read the full conversation, redirect mid-run when the agent drifts, and answer approval prompts without walking back to the desk. That matches how an 82.7 Terminal-Bench-class agent actually gets used — long terminal loops, not single-shot chat.

For the phone-control pattern more generally, see Claude Code from your phone and the adjacent Claude Opus 5 for coding news post (same “configure the agent, SeaWork steers it” shape).

FAQ

What is DeepSeek V4 Flash-0731? The official DeepSeek-V4-Flash checkpoint announced July 31, 2026. API model ID remains deepseek-v4-flash. Per DeepSeek, it is the same architecture as Flash Preview, re-post-trained for stronger agent performance.

Is DeepSeek V4 Pro released? No official V4-Pro GA. A deepseek-v4-pro API has existed since April 2026; DeepSeek says the official Pro release will follow soon, and today’s update does not upgrade Pro.

What is the DeepSeek V4 Flash price? Per DeepSeek’s pricing page: $0.14 per 1M input tokens (cache miss), $0.0028 cache hit, $0.28 per 1M output tokens. Context 1M; max output 384K.

Does Flash-0731 work with Codex / OpenRouter? Codex: yes for Flash, via Responses API — DeepSeek documents the config and a one-click setup script. OpenRouter: third-party routes for V4 Flash (including dated 0731 variants) exist; prices and feature parity vary by provider.

How does 82.7 Terminal-Bench compare to Pro Preview? Per DeepSeek’s HF table: Flash-0731 82.7, V4-Pro-Preview 72.1, Flash Preview 61.8 on Terminal Bench 2.1. Vendor-reported under DeepSeek Harness max effort.

Where to start

1. Read DeepSeek’s 2026-07-31 changelog and pricing.
2. Point Codex at deepseek-v4-flash using DeepSeek’s Codex guide.
3. Optionally self-host weights from DeepSeek-V4-Flash-0731 on Hugging Face (MIT).
4. Leave the agent on the machine that holds the repo; use SeaWork to watch tools, logs, and approvals from your phone.

If you still need Pro-vs-Flash positioning for non-agent bulk work, keep the DeepSeek V4 coding guide open — and treat Flash-0731 as the agent cut that shipped today.

Sources

1. DeepSeek API Change Log — 2026-07-31 DeepSeek-V4-Flash Update — official Flash-0731 public beta, agent benchmark list (Terminal Bench 2.1 82.7), Responses/Codex, Pro “will follow soon,” Flash-only scope.
2. DeepSeek Models & Pricing — deepseek-v4-flash $0.14 / $0.0028 / $0.28, 1M context, 384K max output, Responses API Flash-only footnote, concurrency limits.
3. Hugging Face: deepseek-ai/DeepSeek-V4-Flash-0731 — MIT license, full agent comparison table (Flash-0731 vs Flash Preview vs V4-Pro-Preview), eval notes, lastModified 2026-07-31.
4. DeepSeek Codex integration — Flash-only Codex support today; Pro expected early August 2026.
5. DeepSeek V4 model card PDF — Flash 285B total / 13B active, 1M context, MIT distribution for open-source weights (April 2026 card; architecture size unchanged for 0731 per DeepSeek).
6. OpenRouter: DeepSeek V4 Flash — community/third-party routing listing (~284B total / 13B active wording on that page; provider prices differ from official API).
7. TechNode: DeepSeek puts V4-Flash API into public beta (2026-07-31) — secondary confirmation of public beta, 82.7 Terminal Bench, Responses/Codex, Pro API unchanged.